AI Eats the Grunt Work. It Also Eats the Path to Becoming an Expert.
The formative tasks firms are handing to models are the same ones that used to manufacture senior judgement. Automate the apprenticeship and the seniority bench empties on a delay nobody budgeted for, leaving fluent output and no one left who was trained to check it.
Every organisation that runs on expert judgement bought that judgement second-hand. Think of the analyst who smells a bad deal in a spreadsheet before the model has finished loading, or the engineer who names the failure mode before the demo falls over. Neither of them arrived able to do it. They were manufactured slowly, by years of unglamorous work: reading the filings nobody wanted to read, drafting the memo that got torn up, chasing a number that would not reconcile, producing the first pass a senior then bled red ink all over.
That unglamorous work is exactly what this wave of AI is good at. Today's experts keep their edge. Tomorrow's never get made. That is the bill nobody has priced, and it comes due on a delay.
Where does expertise actually come from?
Follow the mechanism. Expertise gets compiled slowly, one low-value task at a time, each task leaving a residue of judgement behind. The junior does the grind for a reason that has nothing to do with the grind itself: doing it is how the pattern recognition gets installed. One recent analysis calls this the depletion of judgement capital, arguing that even lawful, carefully governed AI use can normalise outsourcing the very activities through which judgement is formed. That is an argument rather than a measurement, but the mechanism it describes is hard to wave away.
Take one concrete task. A junior equity analyst used to build the comparable-company table by hand: pulling each rival's accounts, reconciling the reported figures against the notes, stripping out the one-off gains management had folded into 'adjusted' earnings. The table itself was mostly thrown away. What it built, invisibly, was a nose for which adjustments are legitimate and which are a company flattering itself, an instinct that only forms after you have reconciled a few hundred of them and been fooled by a dozen. Hand that job to a model and the table arrives in seconds, correctly formatted, every add-back included exactly as management reported it. The junior keeps the output and never grows the nose. Five years on, when a valuation turns on spotting an aggressive add-back, the model still produces the table and nobody in the room has the instinct to challenge it.
Break that loop and the arithmetic changes quietly. You still get the deliverable. A model produces the market overview faster and cheaper than any graduate, which is why firms are routing precisely that work to it. What you stop producing is the graduate who, five years on, could have looked at the model's overview and known which line was wrong.
Why firms can't see the bill coming
The erosion is invisible because it sits off the balance sheet and off the clock. Cutting the junior research task shows up this quarter as a cost saving and a speed gain, and both are real. The liability it creates, an expertise pipeline running dry, shows up years later as a hiring problem with no obvious cause. The incentive trap is textbook: the saving is legible and immediate, the damage is diffuse and deferred, so rational managers optimise straight into it. What sharpens the trap is that the immediate payoff is often unproven too. Accenture's own reporting describes AI value realisation falling short of expectations even as the firm plans job reductions, which is the incentive at its starkest: the headcount saving gets booked now, against a return that has yet to show up.
Be clear about the status of the claim. Nobody has a dataset showing a seniority bench collapsing on a fixed schedule, because the effect is slow, diffuse and years from maturing. What follows is a projection built from a mechanism, not a measured trajectory. By the time the current seniors retire and nobody in the building can replace them, the decisions that emptied the bench will be long forgotten and unattributable, which is exactly why the argument has to be made before the evidence is conclusive rather than after.
The verification problem eats itself
Here is the second-order consequence that should worry any technical leader. A model's output still needs a human who can check it, and checking is a senior skill: the compiled judgement to look at a fluent, confident answer and see that it is subtly wrong. That skill was built by doing the work the model now does. So the capability you most need in an AI-heavy operation, someone who can supervise the machine, is the capability whose training ground you just automated away. You spend down your stock of verifiers while closing the factory that makes them, and it converges on one place: fluent output nobody in the room is qualified to challenge.
How hard is the checking, really? Harder than the people doing it tend to think. When METR ran a randomised trial with experienced open-source developers in 2025, the AI assistance made them 19% slower even as they believed it was making them 20% faster. One study, one setting, and worth no more than that. But it lands on the exact nerve: if seasoned practitioners cannot reliably judge whether the tool helped their own work, expecting an under-trained junior to catch where its output is wrong is optimistic to the point of negligence.
The honest rebuttal is that apprenticeship can be redesigned rather than mourned. Maybe juniors learn faster by interrogating a model's reasoning than by grinding out first drafts. Maybe judgement can be taught by putting people in the editor's seat from day one. It is plausible, and it might even be better. But note what it requires: a deliberate act of design. Nobody arrives there by default. The default path hands juniors the answer instead of the search, and trains them to trust output they cannot evaluate. Speed-running the apprenticeship is possible. Deleting it by accident is what is actually happening.
What to do before the bench empties
None of this argues against the tools. It argues for treating expertise as infrastructure you have to maintain on purpose. Three moves follow from the mechanism. First, decide which formative work is training and which is merely cost, and protect the former even when a model could do it cheaper. Second, rebuild the junior job around verification and challenge rather than drafting, so people earn judgement by pressure-testing machine output instead of being replaced by it. Third, measure the thing that is actually decaying: not headcount or output, but how many people in the building can reliably catch a wrong answer.
Doing that well is a strategy question before it is a tooling question, which is why the sensible starting point is a clear-eyed look at where human judgement genuinely adds value in your workflow, the kind of assessment that belongs in any honest AI-readiness review rather than in the rollout afterthoughts. It also argues for designs that keep a competent human in the loop by construction, not as a compliance box. Get the operating model right and the tools compound your people; get it wrong and they hollow them out. If you are weighing that trade for real, it is the core of what a technical strategy engagement exists to answer.
So put the question to a board in concrete terms. A decade from now, when a valuation or a safety case or a contract hinges on catching the model in a confident mistake, who on the payroll will still have the instinct to do it? That capacity is being spent today, quietly, by managers who were only ever asked to hit this quarter's number and had no line on the form for the experts they were deciding not to build.
Questions people ask
Will AI make junior knowledge workers redundant?
Their job changes shape rather than disappearing. The drafting and aggregation that used to fill a junior's day is being automated, so the role has to be rebuilt around verifying, challenging and improving machine output. Firms that don't redesign it deliberately end up with juniors who never build the judgement to supervise the tools they depend on.
How do you train expertise when AI does the entry-level work?
Move people to the editor's seat earlier. Instead of having juniors produce first drafts, have them interrogate a model's reasoning, find its errors and defend their corrections. That can build judgement faster than the old grind, but only if it's designed on purpose. Left to default, teams hand juniors the answer and train them to trust output they can't evaluate.
What is the apprenticeship gap in AI adoption?
It's the delayed shortage of senior judgement created when firms automate the formative, low-value work that used to turn juniors into experts. You keep the output but stop producing the next generation of seniors. The gap is invisible for years, then surfaces as a hiring crisis when the current senior bench retires with no qualified replacements.
Related
- AI Deskilling Is the Workforce Risk Your Board Isn't Pricing
- AI Therapy's Business Case Is a Floor, Not a Ceiling
- The Memory Oligopoly Behind AI's Cost Inflation
- AI & Automation
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.